(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No. 7, 2014
Breast Cancer Diagnosis using Artificial Neural
Networks with Extreme Learning Techniques
Chandra Prasetyo Utomo
YARSI E-Health Research Center YARSI E-Health Research Center
Faculty of Information Technology Faculty of Information Technology
YARSI University, YARSI University,
Jakarta, Indonesia
Abstract—Breast cancer is the second cause of dead among
women. Early detection followed by appropriate cancer
treatment can reduce the deadly risk. Medical professionals can
make mistakes while identifying a disease. The help of technology
such as data mining and machine learning can substantially
improve the diagnosis accuracy. Artificial Neural Networks
(ANN) has been widely used in intelligent breast cancer
diagnosis. However, the standard Gradient-Based Back
Propagation Artificial Neural Networks (BP ANN) has some
limitations. There are parameters to be set in the beginning, long
time for training process, and possibility to be trapped in local
minima. In this research, we implemented ANN with extreme
learning techniques for diagnosing breast cancer based on Breast
Cancer Wisconsin Dataset. Results showed that Extreme
Learning Machine Neural Networks (ELM ANN) has better
generalization classifier model than BP ANN. The development of
this technique is promising as intelligent component in medical
decision support systems.
Keywords—breast cancer; artificial neural networks; extreme
learning machine; medical decision support systems
I. INTRODUCTION
The out of control development of cells in an organ is
called tumors that can be cancerous. There are two kinds of
tumors, benign and malignant. Benign or non-cancerous tumors
are not spreading and are not life intimidating. In the other
hand, malignant or cancerous tumor are expanding and life
threatening [1]. Malignant breast cancer is defined when the
growing cells are in the breast tissue. Breast cancer is the
second overall cause of mortality among women and the first
cause of dead among them between 40 and 55 ages [2].
Regular breast cancer diagnosis followed by appropriate
cancer treatment can reduce the unwilling risk. It is suggested
to do tumor evaluation test every 4-6 weeks. Based on that
reason, benign and malignant detection based on classification
features become very important [3].
Careful diagnosis in early detection has been proven to
lessen the dead rate because of breast cancer [4]. Depend on the
expertise, mistakes can be made by medical professionals while
identifying a disease. With the help of technology such as data
mining and machine learning, diagnosis can be more accurate
(91.1%) when related to a diagnosis made by an experienced
doctor (79.9%) [5].
This research is funded by YARSI University in the Internal Research
Grant Schema 2013/2014.
Aan Kardiana Rika Yuliwulandari
YARSI Genetics Research Center
Faculty of Medicine
YARSI University, Jakarta,
Jakarta, Indonesia Indonesia
ANN is one of the best artificial intelligence techniques for
common data mining tasks, such classification and regression
problems. A lot of research showed that ANN delivered good
accuracy in breast cancer diagnosis. However, this method has
several limitations. First, ANN has some parameters to be
tuned in the beginning of training process such as number of
hidden layer and hidden nodes, learning rates, and activation
function. Second, it takes long time for training process due to
complex architecture and parameters update process in each
iteration that need expensive computational cost. Third, it can
be trapped to local minima so that the optimal performance
cannot be guaranteed. Numerous efforts had been attempted to
get the solutions of neural networks limitations. Huang and
Babri [6] proved that Single Hidden Layer Neural Networks
(SFLN) with tree steps extreme learning process that called
ELM can solve that problems.
In this paper, we revealed the implementation of artificial
neural networks with extreme learning techniques in breast
cancer diagnosis. The dataset used for experiments was Breast
Cancer Wisconsin Dataset that was obtained from the
University of Wisconsin Hospital, Madison from Dr. William
H. Wolberg [7]. We compared the perfomance of ELM with
conventional BP ANN with gradient descent based learning
algorithms. Sensitivity, specificity, and accuracy were used as
performance measurements. Results showed that ELM ANN
generally produced better result than BP ANN.
The rest of this paper is organized as the following. Section
2 is dedicated as literature review. In this section, brief review
of previous works in breast cancer diagnosis are presented. In
Section 3, the concept, mathematical model, and training
process of extreme learning machine are explained. In Section
4, experiments, results, and analysis are provided. Finally,
conclusions and future works are given in Section 5.
II. LITREATURE REVIEW
The uses of classification systems in medical diagnosis,
including breast cancer diagnosis, are growing rapidly.
Evaluation and decision making process from expert medical
diagnosis is key important factor. However, intelligent
classification algorithm may help doctor especially in
minimizing error from unexperienced practitioners [3].
Several techniques have been deployed to predict and
recognize meaningful pattern for breast cancer diagnosis. Ryua
10 | P a g e
www.ijarai.thesai.org
 (IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No. 7, 2014

[8] developed data classification method, called isotonic βi = [βi1, βi2, . . . , βip]T,
separation. The performances were compared against support (3)
vector machines, learning vector quantization, decision tree
induction, and other methods based on two-breast cancer data wi ∙ x(k) is the inner product of wi and x(k),
set, sufficient and insufficient data. The experiment results bi is the bias of the i-th hidden node,
demonstrated that isotonic separation was a practical tool for o(k) ϵ Rp is the output of neural network for k-th vector.
classification in the medical domain.
Hybrid machine learning method was applied by Sahan [9] SLFN can approximate m vectors means that there are exist
in diagnosing breast cancer. The method hybridized a fuzzy- wi, βi, and bi, such that:
artificial immune system with k-nearest neighbour algorithm.
The hybrid method delivered good accuray in Wisconsin
Breast Cancer Dataset (WBCD). They believe it can also be
(4)
tested in other breast cancer diagnosis problems.
Comprehensive view of automated diagnostic systems
implementation for breast cancer detection was provided by
Ubeyli [10]. It compared the performances of multilayer
perceptron neural network (MLPNN), combined neural
network (CNN), probabilistic neural network (PNN), recurrent (5)
neural network (RNN) and support vector machine (SVM). The
aim of that works was to be a guide for a reader who wants to
develop this kind of systems. Equation (5) can be written as:
Numerous combinations and hybrid systems used neural
networks as a component. However, since almost all of the , (6)
employed neural networks are conventional gradient descent
BP ANN, the novel or hybrid method still suffered the neural where
networks drawbacks that were mentioned in the previous
section. H є Rm x M is the hidden layer output matrix of the neural
III. EXTREME LEARNING MACHINE networks.
Huge efforts had been attempted to solve the weaknesses of
 g ( w1  x (1)  b1 )  g ( wM  x (1)  bM ) 
BP ANN. Huang and Babri [6] demonstrated that single hidden H=   (7)
layer feedforward neural networks (SLFN) with at most m     
hidden nodes was capable to estimate function for m different  g ( w1  x ( m )  b1 )  g ( wM  x ( m )  bM )
 
vectors in training dataset.
Given m instances in D = {(x(k), t(k)) | x(k) ϵ Rn, t(k) ϵ Rp, k = β є RM xp
is the weights between hidden and output layers
1,..,m} as training dataset where x(k) = [x1(k), x2(k), …., xn(k)]T as
features and t(k) = [t1(k), t2(k), …., tp(k)]T as target. A SLFN with M
  1T 
number of hidden nodes, activation function g(x) in hidden  
β=   , (8)
nodes, and linear activation function in output nodes is
 T 
mathematically wrote as:
 M 

T є Rm x p is the target values of m vectors in training dataset

 t (1) T 
(1)  
T=   , (9)
where t ( m ) T 
 
wi ϵ Rn is the weights between the input nodes and the i-th
hidden node In the traditional gradient descent based learning algorithm,
weights wi which was connecting the input layer and hidden
wi = [wi1, wi1, . . . , win ]T, layer and biases bi in the hidden nodes were needed to be
(2) initialized and tuned in every iteration. This was the main
factor which often made training process of neural became time
βi ϵ Rp is the weights between the i-th hidden node and the consuming and the trained model may not reach global
minima.
output nodes

11 | P a g e
www.ijarai.thesai.org
 (IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No. 7, 2014

Huang [11] proposed minimum norm least-squares solution Cancer Wisconsin Dataset obtained from the University of
of SLFN which didn’t need to tune those parameters. Training Wisconsin Hospital, Madison from Dr. William H. Wolberg
SLFN with fixed input weights wi and the hidden layer biases bi [7]. The data has 699 instances with 10 attributes plus the class
was similar to find a least square solution of the linear attributes. The class distribution are 65.5% (458 instances) for
benign and 34.5% (241 instances) for malignant. The attribute
system :
information can be seen in TABLE I.
– TABLE I. ATTRIBUTE INFORMATION
(10) # Attributes Domain
1 Sample Code Number id number
The smallest norm least squares solution of that linear 2 Clump Thickness 1 – 10
system was 3 Uniformity of Cell Size 1 – 10
4 Uniformity of Cell Shape 1 – 10
(11) 5 Marginal Adhesion 1 – 10
6 Single Epithelial Cell Size 1 – 10
7 Bare Nuclei 1 – 10
where was the Moore-Penrose generalized inverse of 8 Bland Chromatin 1 – 10
matrix H. This solution had three important properties which 9 Normal Nucleoli 1 – 10
were minimum training error, smallest norm of weights, and 10 Mitoses 1 – 10
unique solution which is . 11 Class 2: benign; 4: malignant

The above minimum norm least-square solution for SLFN

was called extreme learning machine (ELM). Given m
instances in training dataset D = {(x(k), t(k)) | x(k) є Rn, t(k) є Rp, k
= 1,..,m}, activation function g(x), and number of hidden node
M. The training process of ELM is the the following:
1. Randomly set input-hidden layer weights wi and bias
bi, i = 1,…,M.
2. Compute the matrix of hidden layer output H
3. Compute the hidden-output layer weights for
where T = [t(1),…, t(m)].
Based on that definition, there are three main differences
between BP ANN and ELM ANN. First, BP ANN needs to
tuning several parameters, such as number of hidden nodes,
learning rates, momentum, and termination criteria. On the
other hand, ELM ANN is a simple tuning free algorithm. The
only one to be defined is number of hidden nodes. Second, BP
ANN works only for differentiable activation functions in
hidden and output nodes while ELM ANN can use both
differentiable and undifferentiable activation functions.
Finally, BP ANN get trained model which has minimum
training error so that there is a possibility to finish in local
minima. On the other hand, ELM ANN get trained model
which has minimum training error and smallest norm of weight
so that it can produce better generalization model and reach
global minima [12].
IV. EXPERIMENTS, RESULTS, AND ANALYSIS
This section discussed about experimental design,
generated results, and analytical process in order to get valid
conclusion.
A. Experiments
The experiments consisted of three main steps, which were
data gathering, data preprocessing, and performance
evaluating. The dataset used in this experiment was Breast
In the second step, the raw dataset was preprocessed to
produce well-from data that suitable for training and testing
process. The first attribute, sample code number, was removed
because it was not relevant to the diagnosis. The next nine
attributes were normalized into [-1, 1] and used as predictor.
The last attribute was transformed to 0 (benign) and 1
(malignant) such that it can be properly fitted to the standard
BP ANN and ELM ANN implementation.
The method in this experiment was k-fold crossvalidation
with k = 5. This means, the data were randomly divided into 5
partitions. There were 5 experiments. In the each experiment, a
partition was used as testing data and the rest partitions were
treated as training data.
The standard performance measurement for classification
problem was accuracy. However, since the class distribution
was not balanced, it was important to use specificity and
sensitivity as supplementary measurements. In addition, to
minimize the effect of random generated weights in BP ANN
and ELM ANN, each experiment was run three times and the
average results were noted.
(13)
(14)
(15)
B. Results and Analysis
With 5-fold crossvalidation method and each experiment
were run three times, there will be 15 experiments in total. The
whole steps had been done in computer with Intel® Core™ i3,
4096MB RAM, and Windows 7 OS. The results of ELM are
given in TABEL II.

12 | P a g e
www.ijarai.thesai.org
 (IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No. 7, 2014

TABLE II. ELM PERFORMANCES

1.000

Average Sensitivity Rates

Experiments Running Sensitivity Specificity Accuracy
1 0,902 0,965 0,942 0.800
#1 2 0,882 0,977 0,942 0.600
3 0,882 0,965 0,934
1 0,975 0,958 0,963 0.400 BP ANN
#2 2 0,950 0,979 0,971 ELM ANN
0.200
3 1,000 0,969 0,978
1 0,909 0,988 0,956 0.000
#3 2 0,964 0,963 0,963 1 2 3 4 5
3 0,964 0,975 0,971 Experiments #
1 0,957 1,000 0,985
#4 2 0,957 0,989 0,978 Fig. 1. Average sensitivity rates of BP ANN dan ELM
3 0,957 1,000 0,985
1 0,957 0,967 0,964 1.000

Average Specificity Rates

#5 2 1,000 0,956 0,971
3 0,957 0,956 0,956 0.800
In order to compare the ELM ANN performances, training 0.600
and testing with BP ANN were conducted with identical
experimental design. The performances of BP ANN can be 0.400 BP ANN
seen in TABLE III. ELM ANN
0.200

TABLE III. BP ANN PERFORMANCES 0.000

Experiments Running Sensitivity Specificity Accuracy 1 2 3 4 5
1 0,875 0,975 0,934 Experiments #
#1 2 0,750 0,988 0,891
3 0,804 0,963 0,898 Fig. 2. Average specificity rates of BP ANN dan ELM
1 0,855 0,987 0,927
#2 2 0,871 0,973 0,927
3 0,903 0,973 0,941
1.000
Average Accuracy Rates

1 0,825 0,949 0,897 0.800

#3 2 0,860 0,987 0,934
3 0,825 0,975 0,912 0.600
1 0,790 0,975 0,898 BP ANN
0.400
#4 2 0,825 0,975 0,912
ELM ANN
3 0,807 0,975 0,905 0.200
1 0,892 1,000 0,949
#5 2 0,892 1,000 0,949 0.000
1 2 3 4 5
3 0,877 1,000 0,942
Experiments #
To have clear view between the performances of ELM
ANN and BP ANN, the results were transformed to graphical
charts. In each performance measurement, the average values Fig. 3. Average accuracy rates of BP ANN dan ELM
were computed in each experiment. Fig 1 shows the
comparison of average sensitivity rates between BP ANN and Based on Fig 1 and Fig 3, we can see that ELM ANN was
ELM ANN. The comparison of specificity rates were given in superior compared to BP ANN. ELM ANN has better
Fig 2 while accuracy rates can be seen in Fig 3. performances in term of sensitivity and accuracy in all
experiments.

13 | P a g e
www.ijarai.thesai.org
 (IJARAI) International Journal of Advanced Research in Artificial Intelligence,
However, in term of specificity, BP ANN has better
performance in three experiments which were experiment #1,
#2, and #5.
To get general conclusion, overall comparison need to be
computed. In each performance measurement, commulative
average rates were matched. Fig 4 displays the whole
sensitivity, specificity, and accuracy average rates between BP
ANN and ELM ANN. Result showed that, generally ELM
ANN were better than BP ANN.
0.980 0.974
1.000 0.948 0.964
0.921
0.900 0.843
0.800
0.700
0.600
0.500
Sensitivity Specificity Accuracy
BP ANN ELM ANN
Fig. 4. Overall average rates between BP ANN dan ELM
V. CONLUSION AND FUTURE WORKS
The performances of ELM ANN were generally better than
BP ANN in breast cancer diagnosis. Although the specificity
rate was slightly lower than BP ANN, it can be clearly seen
that ELM ANN remarkably improved the sensitivity and
accuracy rates. Based on these results, we can conclude that
ELM ANN has better generalization model than BP ANN in
diagnosing breast cancer based on Breast Cancer Wisconsin
Dataset.
There are some necessary works to be done in near future.
First, it is important to communicate the model with domain
expert. The hybrid of ELM ANN with Decision Tree or any
other technique that can produce meaningful knowledge
representation will be promising. Second, to make intelligent
diagnosis tool that can be used by end user, it is necessary to
develop interactive user interface.
www.ijarai.thesai.org
Vol. 3, No. 7, 2014
The development of interfaces in mobile, desktop, or web
application may be useful. Third, there are new cases added
regularly in the hospital. Developing intelligent diagnosis
systems that can not only learn from available data in
repositories but also from newly available data will be required.
REFERENCES
[1] T. S. Subashini, V. Ramalingam, and S. Palanivel, “Breast mass
classification based on cytological patterns using RBFNN and SVM,”
Expert Syst. Appl., vol. 36, no. 3, pp. 5284–5290, Apr. 2009.
[2] TheAmericanCancerSociety, “What is Breast Cancer.” .
[3] M. F. Akay, “Support vector machines combined with feature selection
for breast cancer diagnosis,” Expert Syst. Appl., vol. 36, no. 2, pp. 3240–
3247, Mar. 2009.
[4] D. West, P. Mangiameli, R. Rampal, and V. West, “Ensemble strategies
for a medical diagnostic decision support system: A breast cancer
diagnosis application,” Eur. J. Oper. Res., vol. 162, no. 2, pp. 532–551,
Apr. 2005.
[5] R. W. Brause, “Medical Analysis and Diagnosis by Neural Networks,”
in Proceeding ISMDA ’01 Proceedings of the Second International
Symposium on Medical Data Analysis, 2001, pp. 1–13.
[6] G. B. Huang and H. A. Babri, “Upper bounds on the number of hidden
neurons in feedforward networks with arbitrary bounded nonlinear
activation functions.,” IEEE Trans. Neural Netw., vol. 9, no. 1, pp. 224–
9, Jan. 1998.
[7] W. H. Wolberg and O. L. Mangasarian, “Multisurface method of pattern
separation for medical diagnosis applied to breast cytology,” in
Proceedings of the National Academy of Sciences, U.S.A., Volume 87,
December 1990, pp. 9193–9196.
[8] Y. U. Ryu, R. Chandrasekaran, and V. S. Jacob, “Breast cancer
prediction using the isotonic separation technique,” Eur. J. Oper. Res.,
vol. 181, no. 2, pp. 842–854, Sep. 2007.
[9] S. Sahan, K. Polat, H. Kodaz, and S. Güneş, “A new hybrid method
based on fuzzy-artificial immune system and k-nn algorithm for breast
cancer diagnosis.,” Comput. Biol. Med., vol. 37, no. 3, pp. 415–23, Mar.
2007.
[10] E. D. Übeyli, “Implementing automated diagnostic systems for breast
cancer detection,” Expert Syst. Appl., vol. 33, no. 4, pp. 1054–1062,
Nov. 2007.
[11] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine:
Theory and applications,” Neurocomputing, vol. 70, no. 1–3, pp. 489–
501, Dec. 2006.
[12] C. P. Utomo, “The Hybrid of Classification Tree and Extreme Learning
Machine for Permeability Prediction in Oil Reservoir,” Int. J. Comput.
Sci. Issues, vol. 10, no. 1, pp. 52–60, 2013.
14 | P a g e